Detecting early gastric cancer: Comparison between the diagnostic ability of convolutional neural networks and endoscopists

Abstract

Objectives: Detecting early gastric cancer is difficult, and it may even be overlooked by experienced endoscopists. Recently, artificial intelligence based on deep learning through convolutional neural networks (CNNs) has enabled significant advancements in the field of gastroenterology. However, it remains unclear whether a CNN can outperform endoscopists. In this study, we evaluated whether the performance of a CNN in detecting early gastric cancer is better than that of endoscopists.

Methods: The CNN was constructed using 13,584 endoscopic images from 2639 lesions of gastric cancer. Subsequently, its diagnostic ability was compared to that of 67 endoscopists using an independent test dataset (2940 images from 140 cases).

Results: The average diagnostic time for analyzing 2940 test endoscopic images by the CNN and endoscopists were 45.5 ± 1.8 s and 173.0 ± 66.0 min, respectively. The sensitivity, specificity, and positive and negative predictive values for the CNN were 58.4%, 87.3%, 26.0%, and 96.5%, respectively. These values for the 67 endoscopists were 31.9%, 97.2%, 46.2%, and 94.9%, respectively. The CNN had a significantly higher sensitivity than the endoscopists (by 26.5%; 95% confidence interval, 14.9-32.5%).

Conclusion: The CNN detected more early gastric cancer cases in a shorter time than the endoscopists. The CNN needs further training to achieve higher diagnostic accuracy. However, a diagnostic support tool for gastric cancer using a CNN will be realized in the near future.

Keywords: artificial intelligence; convolutional neural network; deep learning; endoscopy; gastric cancer.

Figure 1

Patient recruitment flowchart.

Figure 2

Representative gastric cancer and non‐cancer endoscopic images. (a) A slightly reddish and depressed lesion of gastric cancer appears on the lesser curvature of the antrum. [0–IIc, 10 mm, tub1, T1a(M)]. (b) This image shows the Helicobacter pylori uninfected gastric mucosa. There is no cancer.

Figure 3

Definition of correct answer. (a) A reddish, depressed lesion of gastric cancer appears on the greater curvature of the lower body. [0–IIc, 9 mm, tub1, T1a(M)]. (b) The correct marking is the red rectangle. The green rectangle is the convolutional neural network (CNN) marking, and the blue rectangle is the endoscopists’ marking. In this case, when the correct marking and the marking of the CNN or endoscopists overlap by 40% or more, they were judged to be correct.

Figure 4

This graph shows the receiver operating characteristic curves for the convolutional neural network (CNN) and predictions of the endoscopists. Each endoscopist's prediction is represented by a single point. The CNN outputs a gastric cancer probability score per image, and the program then calculates a mean square of the probabilities per image. The area under the curve is 75.7%. At a cut‐off value of 0.412, the sensitivity and specificity of the CNN were 58.4% and 87.3, respectively.

Figure 5

Representative images of false positives. The green rectangular frames show areas that the convolutional neural network misdiagnosed as gastric cancer. (a) Spotty redness associated with Helicobacter pylori (H. pylori) infection (gastritis). (b) Cardia (normal anatomical structure). (c) White scar (S2 stage) at the lesser curvature of the upper body (ulcer scar).

Figure 6

Representative images of false negatives. The following cancers were misdiagnosed and the assumed causes for this misdiagnosis were as follows. (a) 0– IIc, 4 mm, tub1, T1a (too small lesion). (b) Images from tangential line (tangential line). (c) Lesion at the angle captured about 7 cm away (too distant lesion).